FluidInference / gliclass-edge-apps-coreml

GLiClass Edge Apps v2 · Core ML

Application-tuned, 32,705,154-parameter GLiClass model converted to Core ML for FluidUse . It receives a state, an optional instruction, and 2–25 candidate labels, then returns a probability for each label in one encoder pass. It does not generate text. The source checkpoint is knowledgator/gliclass-edge-v3.0 at revision df03993a2ed98e5e4a0d2dd7efbbd105abe874cf , further tuned for FluidUse application decisions before conversion. The tokenizer is included.

Files

The .mlpackage folders are portable Core ML models targeting iOS 17 / macOS 14 or newer. The FluidUse loader compiles them on the local machine. L128, L256, and L512 support up to 25 supplied labels; labels must fit inside the selected token limit. The LUT8 package is available only at L128. The Swift runtime chooses the smallest loaded bucket that fits and truncates long state text at its largest loaded bucket; it rejects requests if their option labels no longer fit. Inspect GLiClassAnswer.textWasTruncated when this matters.

Use with FluidUse

Download this repository to a local directory, then load one or more buckets:

hf download FluidInference/gliclass-edge-apps-coreml --local-dir ./gliclass-assets

import FluidUse
import Foundation

let manager = try await GLiClassManager.load(
    from: URL(fileURLWithPath: "./gliclass-assets", isDirectory: true),
    configuration: .init(lengths: [128, 256, 512], precision: "fp16"))
let answer = try await manager.classify(
    text: "The next support request asks for a refund after a duplicate charge.",
    labels: ["refund", "shipping", "technical support"],
    prompt: "Which team should handle this request?")
print(answer.selectedLabel, answer.probabilities)

For the compact bucket, set lengths: [128] and precision: "lut8" . The FluidUse 2048 demo and Tetris demo use GLiClass to compare shortlisted moves. They use game heuristics to prepare candidates; the model makes the final choice among those candidates.

Verification and limits

The converted FP16 model changed 18 argmax choices out of 3,899 application-suite rows relative to its PyTorch checkpoint. On an Apple M5 Pro, the L128 FP16 bucket measured 0.843 ms median model latency; LUT8 measured 0.903 ms and agreed with FP16 on 97.1% of L128 application-suite choices. These are local measurements under the protocol in RESULTS.md , with supporting JSON in reports/ .

A local reproduction of the 19 static Decision Index headline benchmarks scored 19.27. It was not submitted as an official tracker result because the full public bundle was unavailable during the run. The model was tuned on application tasks, not games, and a game result depends on its action shortlist and harness.

Training, conversion, verification, and pinned sources are documented in the FluidInference conversion toolkit . The original GLiClass model is by Knowledgator. This application-tuned Core ML variant is released under Apache-2.0; see LICENSE .